Announcing RingRing Mode

[otakustay]

Thinking

In 2025, we observe a more diverse development trend in community models. High-quality models like Claude, deep reasoning models led by GPT-O1, and high-speed & low-cost models such as DeepSeek V3 have all gained user recognition in their respective fields.

However, when an intelligent coding tool faces such a rich ecosystem of models, we notice that tools like Cursor, Windsurf, and Trae allow users to choose a single model to complete tasks. Personally, I hold a relatively negative view of this approach. In fact, using a single model, we inevitably face a series of issues:

• When using a single highest-quality model, its cost often exceeds what users can afford.
• Using small-scale, high-speed models struggles with complex problems due to their capability limits.
• If we use deep thinking models like DeepSeek R1 to pursue a win-win situation in cost and quality, we encounter the problem of long processing times due to in-depth thinking.

In other words, for developers of coding tools, we are actually facing an "iron triangle" problem. There is no solution that satisfies speed, quality, and cost simultaneously, and any single solution will significantly compromise at least one dimension, leading to a severe decline in user experience.

In response to this dilemma, Oniichan's solution is a hybrid model:

• Let deep thinking models handle the most suitable planning and strategizing tasks.
• Let low-cost models quickly perform information gathering and understanding tasks.
• Let high-quality models generate code and edit files.

Oniichan calls this the RingRing mode, which you can understand as a structure composed of two rings:

• The outer ring continuously generates a plan based on your needs and completes the plan, approaching the final implementation.
• The inner ring completes tasks by using tools according to the tasks specified in the plan.

Through the continuous cycle of these big and small rings, different models perform their respective duties, ultimately completing the entire requirement with high quality.

Demonstration

oniichan-ring-ring-mode.mp4

In the video, you can see a total of 5 plans (conducted by DeepSeek R1), each plan containing multiple tasks, with the execution and completion status of tasks visualized.

Usage

Starting from version 3.4.0, Oniichan's configuration will add 3 new settings:

These 3 model configurations are used to allocate planning, information gathering tasks, and coding tasks. We recommend using the following combination:

• Planning: deepseek/deepseek-r1 with deep thinking, possessing a strong understanding of user needs.
• Information Gathering: deepseek/deepseek-chat with ultra-long context, low cost, and extremely fast speed.
• Coding: anthropic/claude-3.5-sonnet with world-class coding capability.

This combination will allow you to achieve a very good balance between cost and effectiveness. For the same task, compared to using the anthropic/claude-3.5-sonnet single model, the actual cost can be reduced to about 20%.

When initiating a task via the inbox system, you can select the mode in the "To" field. Choose "Oniichan (Ring Ring Mode)" to run this task in RingRing mode.